DeepSeek’s AI Model Faces Major Security Concerns

Chinese AI startup DeepSeek has made headlines with its R1 model, a low-cost, open-source alternative to dominant AI systems from OpenAI and Meta. However, new research has cast serious doubt on its security. A study conducted by Cisco found that DeepSeek R1 failed every single one of 50 adversarial attacks designed to test its safety. This failure rate makes it the most vulnerable mainstream language model tested so far.

Researchers used the HarmBench dataset to assess DeepSeek’s ability to block harmful prompts. Unlike other AI models that at least showed some resistance, DeepSeek R1 reportedly engaged with every request, including those related to cybercrime, misinformation, and illegal activities. Security firm Adversa AI confirmed similar findings, successfully prompting the model to generate instructions on hacking, drug extraction, and other prohibited topics.

The backlash is likely to add to the growing concerns surrounding DeepSeek. Since its release, the model has been scrutinized over its ties to Chinese servers and its ability to censor politically sensitive topics while failing to block harmful content. Critics argue that its safeguards prioritize avoiding state criticism rather than ensuring user safety.

Despite these concerns, DeepSeek continues to gain traction, largely due to its affordability and open-source nature. Unlike competitors reliant on massive infrastructure costs, DeepSeek’s approach promises lower computing expenses, which has made it a disruptive force in the AI landscape. However, these latest security findings may slow its momentum, as researchers and regulators question whether its rapid development came at the cost of responsible AI safeguards.

For now, DeepSeek remains a powerful but deeply flawed model, raising the question of whether cost-effective AI can be achieved without compromising safety.

Source: https://gizmodo.com/deepseek-gets-an-f-in-safety-from-researchers-2000558645?utm_source=tldrai